n8n 3.0 removed the HTTP Request Tool: call Sume from an agent
The legacy HTTP Request Tool is gone in n8n 3.0. Wire the HTTP Request node into the AI Agent Tool input to submit Sume jobs, or use hosted MCP.

On n8n 3.0 the legacy HTTP Request Tool no longer exists, so an agent that called Sume through it needs a replacement. The n8n 3.0 breaking-changes page says to use the regular HTTP Request node connected to the Tool input of the AI Agent node. The other option is to point the agent at Sume's hosted remote MCP server, which exposes the tools directly.
The two replacement paths
Both paths keep the agent in charge of when to generate. They differ in who owns the request shape.
| Path | What you configure | Best when |
|---|---|---|
| HTTP Request node on the Tool input | URL, method, headers and body for one Sume endpoint | You want one fixed call such as a Format run |
| Hosted MCP at https://mcp.sume.com/mcp | Streamable HTTP endpoint plus a bearer API key | You want the agent to discover tools itself |
Path one: a fixed HTTP Request tool
Create the node, set the method to POST and the URL to the Sume endpoint you intend to expose, and add an Authorization: Bearer header from an n8n credential. Add an Idempotency-Key header. Without it, an agent retry after a tool error can create a second paid job.
Describe the tool narrowly. A single Format run endpoint with an output_schema is easier for a model to call correctly than a general generation route. Format runs accept generation_spend_cap_usd, which puts a ceiling on one run.
Path two: hosted MCP
Sume's hosted server is a streamable HTTP endpoint at https://mcp.sume.com/mcp. An API key sent as Authorization: Bearer or x-api-key sees the full tool set. Write and paid tools require an idempotency_key, and accept optional dry_run and max_spend_usd, so the agent can price a call before it commits.
Long jobs matter here too. jobs_wait holds for at most 55 seconds per call. If it returns wait_slice_expired, call it again with the same ids rather than resubmitting the create.
- Use
tools_listandtools_schemafirst so the agent sees real argument names. - A 524 or 522 from a proxy is a transport failure, not a job outcome.
- OAuth tokens are not API keys; pick one auth style per client.
What to test after migrating
Run the agent twice with the same request and confirm the second call returns the original job rather than a new one. Then force a tool error and confirm the agent does not loop on paid creates. Finally check that your n8n instance can reach the Sume host, since 3.0 also widens the SSRF blocklist to include 100.64.0.0/10.
Sources
Related posts
More in Integrations
- n8n 3.0 SSRF blocklist adds 100.64.0.0/10: Sume webhook paths
n8n 3.0 adds 100.64.0.0/10 to its SSRF blocklist. Sume is public HTTPS, so calls are fine, but webhook URLs must be public for Sume to reach them.
- n8n 3.0 task runner timeout 60s: Sume polling in Code nodes
n8n 3.0 cuts the task runner timeout from 300 to 60 seconds. Move Sume job polling out of the Code node into webhook, Wait or async patterns.
- n8n Data table as a Sume job ledger: the 200 MiB default limit
Store each Sume job id in an n8n Data table and upsert it when the webhook arrives. The default cap is 200 MiB per instance, and a full table errors inserts.
- n8n Execution Data node: find a run by its Sume job id
Save the Sume job id with n8n's Execution Data node and you can search the Executions list by it. Keys cap at 50 characters, values at 512, plan limits apply.
Written by Sume